基于阈值图像分割结合深度学习的血细胞识别算法研究
Research of blood cell recognition algorithm based on the combination of threshold image segmentation and deep learning
蔡润秋 1巫琦 1马靖武 1张译1
作者信息
- 1. 南京中医药大学附属医院设备处 南京 210000
- 折叠
摘要
目的:探讨传统图像处理的阈值图像分割与深度学习相结合的血细胞识别算法,用于血细胞涂片的自动识别和分类.方法:使用全局阈值分割从血细胞涂片中提取血细胞并单独保存,对分割后的细胞图像进行人工标记和分类,以降低后续处理的硬件要求.标记后的图像深度学习训练基于GoogLeNet预训练模型,生成可用于血细胞涂片图像自动识别的深度学习模型.使用已训练模型对测试集进行评价,并生成混淆矩阵和受试者工作特征(ROC)曲线下面积(AUC)值.结果:这种预处理被证明可以提高深度学习模型的训练以及后续使用模型识别的速度10倍以上.使用在线血细胞涂片图像数据集Raabin-WBC Data,模型训练的准确率达到93.06%,均取得了良好的结果.结论:基于阈值图像分割与深度学习相结合的血细胞识别算法,可极大提高血细胞识别分类的效率,保证血液相关疾病诊断的准确度.
Abstract
Objective:To explore a blood cell recognition algorithm that combined threshold image segmentation with deep learning in conventional image processing,so as to be used in automatic recognition and classification of blood cell smears.Method:Global threshold segmentation was used to extract blood cells from blood cell smears and to store them separately.The segmented cell images were manually labeled and classified so as to reduce the requirements for hardware in subsequent processing.The deep learning training of labeled images was on the basis of the GoogLeNet pre training model,which could generate deep learning model of automatic recognition that could be used in the images of blood cell smear.The trained model could be used to evaluate the test set,and generate confusion matrix and area under curve(AUC)value of receiver operating characteristic(ROC)curve.Result:This preprocessing has been proven that it can improve the training of deep learning model,and the subsequent recognition speed of using model can exceed over 10 times.Using the online image dataset Raabin WBC Data of blood cell smear,the accuracy of model training reached to 93.06%.Both of them obtained favorable results.Conclusion:The blood cell recognition algorithm based on the combination of threshold image segmentation and deep learning can greatly improve the efficiency of recognition and classification of blood cells,and ensure accuracy of the diagnosis of blood related diseases.
关键词
血细胞识别/阈值分割/深度学习/预训练模型Key words
Blood cells recognition/Threshold segmentation/Deep learning/Pre-training model引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据